Call the filesystem back whenever a page is removed from the page cache
[linux-2.6.git] / mm / vmscan.c
1 /*
2  *  linux/mm/vmscan.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  *
6  *  Swap reorganised 29.12.95, Stephen Tweedie.
7  *  kswapd added: 7.1.96  sct
8  *  Removed kswapd_ctl limits, and swap out as many pages as needed
9  *  to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10  *  Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11  *  Multiqueue VM started 5.8.00, Rik van Riel.
12  */
13
14 #include <linux/mm.h>
15 #include <linux/module.h>
16 #include <linux/gfp.h>
17 #include <linux/kernel_stat.h>
18 #include <linux/swap.h>
19 #include <linux/pagemap.h>
20 #include <linux/init.h>
21 #include <linux/highmem.h>
22 #include <linux/vmstat.h>
23 #include <linux/file.h>
24 #include <linux/writeback.h>
25 #include <linux/blkdev.h>
26 #include <linux/buffer_head.h>  /* for try_to_release_page(),
27                                         buffer_heads_over_limit */
28 #include <linux/mm_inline.h>
29 #include <linux/pagevec.h>
30 #include <linux/backing-dev.h>
31 #include <linux/rmap.h>
32 #include <linux/topology.h>
33 #include <linux/cpu.h>
34 #include <linux/cpuset.h>
35 #include <linux/notifier.h>
36 #include <linux/rwsem.h>
37 #include <linux/delay.h>
38 #include <linux/kthread.h>
39 #include <linux/freezer.h>
40 #include <linux/memcontrol.h>
41 #include <linux/delayacct.h>
42 #include <linux/sysctl.h>
43
44 #include <asm/tlbflush.h>
45 #include <asm/div64.h>
46
47 #include <linux/swapops.h>
48
49 #include "internal.h"
50
51 #define CREATE_TRACE_POINTS
52 #include <trace/events/vmscan.h>
53
54 enum lumpy_mode {
55         LUMPY_MODE_NONE,
56         LUMPY_MODE_ASYNC,
57         LUMPY_MODE_SYNC,
58 };
59
60 struct scan_control {
61         /* Incremented by the number of inactive pages that were scanned */
62         unsigned long nr_scanned;
63
64         /* Number of pages freed so far during a call to shrink_zones() */
65         unsigned long nr_reclaimed;
66
67         /* How many pages shrink_list() should reclaim */
68         unsigned long nr_to_reclaim;
69
70         unsigned long hibernation_mode;
71
72         /* This context's GFP mask */
73         gfp_t gfp_mask;
74
75         int may_writepage;
76
77         /* Can mapped pages be reclaimed? */
78         int may_unmap;
79
80         /* Can pages be swapped as part of reclaim? */
81         int may_swap;
82
83         int swappiness;
84
85         int order;
86
87         /*
88          * Intend to reclaim enough continuous memory rather than reclaim
89          * enough amount of memory. i.e, mode for high order allocation.
90          */
91         enum lumpy_mode lumpy_reclaim_mode;
92
93         /* Which cgroup do we reclaim from */
94         struct mem_cgroup *mem_cgroup;
95
96         /*
97          * Nodemask of nodes allowed by the caller. If NULL, all nodes
98          * are scanned.
99          */
100         nodemask_t      *nodemask;
101 };
102
103 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
104
105 #ifdef ARCH_HAS_PREFETCH
106 #define prefetch_prev_lru_page(_page, _base, _field)                    \
107         do {                                                            \
108                 if ((_page)->lru.prev != _base) {                       \
109                         struct page *prev;                              \
110                                                                         \
111                         prev = lru_to_page(&(_page->lru));              \
112                         prefetch(&prev->_field);                        \
113                 }                                                       \
114         } while (0)
115 #else
116 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
117 #endif
118
119 #ifdef ARCH_HAS_PREFETCHW
120 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
121         do {                                                            \
122                 if ((_page)->lru.prev != _base) {                       \
123                         struct page *prev;                              \
124                                                                         \
125                         prev = lru_to_page(&(_page->lru));              \
126                         prefetchw(&prev->_field);                       \
127                 }                                                       \
128         } while (0)
129 #else
130 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
131 #endif
132
133 /*
134  * From 0 .. 100.  Higher means more swappy.
135  */
136 int vm_swappiness = 60;
137 long vm_total_pages;    /* The total number of pages which the VM controls */
138
139 static LIST_HEAD(shrinker_list);
140 static DECLARE_RWSEM(shrinker_rwsem);
141
142 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
143 #define scanning_global_lru(sc) (!(sc)->mem_cgroup)
144 #else
145 #define scanning_global_lru(sc) (1)
146 #endif
147
148 static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
149                                                   struct scan_control *sc)
150 {
151         if (!scanning_global_lru(sc))
152                 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
153
154         return &zone->reclaim_stat;
155 }
156
157 static unsigned long zone_nr_lru_pages(struct zone *zone,
158                                 struct scan_control *sc, enum lru_list lru)
159 {
160         if (!scanning_global_lru(sc))
161                 return mem_cgroup_zone_nr_pages(sc->mem_cgroup, zone, lru);
162
163         return zone_page_state(zone, NR_LRU_BASE + lru);
164 }
165
166
167 /*
168  * Add a shrinker callback to be called from the vm
169  */
170 void register_shrinker(struct shrinker *shrinker)
171 {
172         shrinker->nr = 0;
173         down_write(&shrinker_rwsem);
174         list_add_tail(&shrinker->list, &shrinker_list);
175         up_write(&shrinker_rwsem);
176 }
177 EXPORT_SYMBOL(register_shrinker);
178
179 /*
180  * Remove one
181  */
182 void unregister_shrinker(struct shrinker *shrinker)
183 {
184         down_write(&shrinker_rwsem);
185         list_del(&shrinker->list);
186         up_write(&shrinker_rwsem);
187 }
188 EXPORT_SYMBOL(unregister_shrinker);
189
190 #define SHRINK_BATCH 128
191 /*
192  * Call the shrink functions to age shrinkable caches
193  *
194  * Here we assume it costs one seek to replace a lru page and that it also
195  * takes a seek to recreate a cache object.  With this in mind we age equal
196  * percentages of the lru and ageable caches.  This should balance the seeks
197  * generated by these structures.
198  *
199  * If the vm encountered mapped pages on the LRU it increase the pressure on
200  * slab to avoid swapping.
201  *
202  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
203  *
204  * `lru_pages' represents the number of on-LRU pages in all the zones which
205  * are eligible for the caller's allocation attempt.  It is used for balancing
206  * slab reclaim versus page reclaim.
207  *
208  * Returns the number of slab objects which we shrunk.
209  */
210 unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
211                         unsigned long lru_pages)
212 {
213         struct shrinker *shrinker;
214         unsigned long ret = 0;
215
216         if (scanned == 0)
217                 scanned = SWAP_CLUSTER_MAX;
218
219         if (!down_read_trylock(&shrinker_rwsem))
220                 return 1;       /* Assume we'll be able to shrink next time */
221
222         list_for_each_entry(shrinker, &shrinker_list, list) {
223                 unsigned long long delta;
224                 unsigned long total_scan;
225                 unsigned long max_pass;
226
227                 max_pass = (*shrinker->shrink)(shrinker, 0, gfp_mask);
228                 delta = (4 * scanned) / shrinker->seeks;
229                 delta *= max_pass;
230                 do_div(delta, lru_pages + 1);
231                 shrinker->nr += delta;
232                 if (shrinker->nr < 0) {
233                         printk(KERN_ERR "shrink_slab: %pF negative objects to "
234                                "delete nr=%ld\n",
235                                shrinker->shrink, shrinker->nr);
236                         shrinker->nr = max_pass;
237                 }
238
239                 /*
240                  * Avoid risking looping forever due to too large nr value:
241                  * never try to free more than twice the estimate number of
242                  * freeable entries.
243                  */
244                 if (shrinker->nr > max_pass * 2)
245                         shrinker->nr = max_pass * 2;
246
247                 total_scan = shrinker->nr;
248                 shrinker->nr = 0;
249
250                 while (total_scan >= SHRINK_BATCH) {
251                         long this_scan = SHRINK_BATCH;
252                         int shrink_ret;
253                         int nr_before;
254
255                         nr_before = (*shrinker->shrink)(shrinker, 0, gfp_mask);
256                         shrink_ret = (*shrinker->shrink)(shrinker, this_scan,
257                                                                 gfp_mask);
258                         if (shrink_ret == -1)
259                                 break;
260                         if (shrink_ret < nr_before)
261                                 ret += nr_before - shrink_ret;
262                         count_vm_events(SLABS_SCANNED, this_scan);
263                         total_scan -= this_scan;
264
265                         cond_resched();
266                 }
267
268                 shrinker->nr += total_scan;
269         }
270         up_read(&shrinker_rwsem);
271         return ret;
272 }
273
274 static void set_lumpy_reclaim_mode(int priority, struct scan_control *sc,
275                                    bool sync)
276 {
277         enum lumpy_mode mode = sync ? LUMPY_MODE_SYNC : LUMPY_MODE_ASYNC;
278
279         /*
280          * Some reclaim have alredy been failed. No worth to try synchronous
281          * lumpy reclaim.
282          */
283         if (sync && sc->lumpy_reclaim_mode == LUMPY_MODE_NONE)
284                 return;
285
286         /*
287          * If we need a large contiguous chunk of memory, or have
288          * trouble getting a small set of contiguous pages, we
289          * will reclaim both active and inactive pages.
290          */
291         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
292                 sc->lumpy_reclaim_mode = mode;
293         else if (sc->order && priority < DEF_PRIORITY - 2)
294                 sc->lumpy_reclaim_mode = mode;
295         else
296                 sc->lumpy_reclaim_mode = LUMPY_MODE_NONE;
297 }
298
299 static void disable_lumpy_reclaim_mode(struct scan_control *sc)
300 {
301         sc->lumpy_reclaim_mode = LUMPY_MODE_NONE;
302 }
303
304 static inline int is_page_cache_freeable(struct page *page)
305 {
306         /*
307          * A freeable page cache page is referenced only by the caller
308          * that isolated the page, the page cache radix tree and
309          * optional buffer heads at page->private.
310          */
311         return page_count(page) - page_has_private(page) == 2;
312 }
313
314 static int may_write_to_queue(struct backing_dev_info *bdi,
315                               struct scan_control *sc)
316 {
317         if (current->flags & PF_SWAPWRITE)
318                 return 1;
319         if (!bdi_write_congested(bdi))
320                 return 1;
321         if (bdi == current->backing_dev_info)
322                 return 1;
323
324         /* lumpy reclaim for hugepage often need a lot of write */
325         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
326                 return 1;
327         return 0;
328 }
329
330 /*
331  * We detected a synchronous write error writing a page out.  Probably
332  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
333  * fsync(), msync() or close().
334  *
335  * The tricky part is that after writepage we cannot touch the mapping: nothing
336  * prevents it from being freed up.  But we have a ref on the page and once
337  * that page is locked, the mapping is pinned.
338  *
339  * We're allowed to run sleeping lock_page() here because we know the caller has
340  * __GFP_FS.
341  */
342 static void handle_write_error(struct address_space *mapping,
343                                 struct page *page, int error)
344 {
345         lock_page_nosync(page);
346         if (page_mapping(page) == mapping)
347                 mapping_set_error(mapping, error);
348         unlock_page(page);
349 }
350
351 /* possible outcome of pageout() */
352 typedef enum {
353         /* failed to write page out, page is locked */
354         PAGE_KEEP,
355         /* move page to the active list, page is locked */
356         PAGE_ACTIVATE,
357         /* page has been sent to the disk successfully, page is unlocked */
358         PAGE_SUCCESS,
359         /* page is clean and locked */
360         PAGE_CLEAN,
361 } pageout_t;
362
363 /*
364  * pageout is called by shrink_page_list() for each dirty page.
365  * Calls ->writepage().
366  */
367 static pageout_t pageout(struct page *page, struct address_space *mapping,
368                          struct scan_control *sc)
369 {
370         /*
371          * If the page is dirty, only perform writeback if that write
372          * will be non-blocking.  To prevent this allocation from being
373          * stalled by pagecache activity.  But note that there may be
374          * stalls if we need to run get_block().  We could test
375          * PagePrivate for that.
376          *
377          * If this process is currently in __generic_file_aio_write() against
378          * this page's queue, we can perform writeback even if that
379          * will block.
380          *
381          * If the page is swapcache, write it back even if that would
382          * block, for some throttling. This happens by accident, because
383          * swap_backing_dev_info is bust: it doesn't reflect the
384          * congestion state of the swapdevs.  Easy to fix, if needed.
385          */
386         if (!is_page_cache_freeable(page))
387                 return PAGE_KEEP;
388         if (!mapping) {
389                 /*
390                  * Some data journaling orphaned pages can have
391                  * page->mapping == NULL while being dirty with clean buffers.
392                  */
393                 if (page_has_private(page)) {
394                         if (try_to_free_buffers(page)) {
395                                 ClearPageDirty(page);
396                                 printk("%s: orphaned page\n", __func__);
397                                 return PAGE_CLEAN;
398                         }
399                 }
400                 return PAGE_KEEP;
401         }
402         if (mapping->a_ops->writepage == NULL)
403                 return PAGE_ACTIVATE;
404         if (!may_write_to_queue(mapping->backing_dev_info, sc))
405                 return PAGE_KEEP;
406
407         if (clear_page_dirty_for_io(page)) {
408                 int res;
409                 struct writeback_control wbc = {
410                         .sync_mode = WB_SYNC_NONE,
411                         .nr_to_write = SWAP_CLUSTER_MAX,
412                         .range_start = 0,
413                         .range_end = LLONG_MAX,
414                         .for_reclaim = 1,
415                 };
416
417                 SetPageReclaim(page);
418                 res = mapping->a_ops->writepage(page, &wbc);
419                 if (res < 0)
420                         handle_write_error(mapping, page, res);
421                 if (res == AOP_WRITEPAGE_ACTIVATE) {
422                         ClearPageReclaim(page);
423                         return PAGE_ACTIVATE;
424                 }
425
426                 /*
427                  * Wait on writeback if requested to. This happens when
428                  * direct reclaiming a large contiguous area and the
429                  * first attempt to free a range of pages fails.
430                  */
431                 if (PageWriteback(page) &&
432                     sc->lumpy_reclaim_mode == LUMPY_MODE_SYNC)
433                         wait_on_page_writeback(page);
434
435                 if (!PageWriteback(page)) {
436                         /* synchronous write or broken a_ops? */
437                         ClearPageReclaim(page);
438                 }
439                 trace_mm_vmscan_writepage(page,
440                         trace_reclaim_flags(page, sc->lumpy_reclaim_mode));
441                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
442                 return PAGE_SUCCESS;
443         }
444
445         return PAGE_CLEAN;
446 }
447
448 /*
449  * Same as remove_mapping, but if the page is removed from the mapping, it
450  * gets returned with a refcount of 0.
451  */
452 static int __remove_mapping(struct address_space *mapping, struct page *page)
453 {
454         BUG_ON(!PageLocked(page));
455         BUG_ON(mapping != page_mapping(page));
456
457         spin_lock_irq(&mapping->tree_lock);
458         /*
459          * The non racy check for a busy page.
460          *
461          * Must be careful with the order of the tests. When someone has
462          * a ref to the page, it may be possible that they dirty it then
463          * drop the reference. So if PageDirty is tested before page_count
464          * here, then the following race may occur:
465          *
466          * get_user_pages(&page);
467          * [user mapping goes away]
468          * write_to(page);
469          *                              !PageDirty(page)    [good]
470          * SetPageDirty(page);
471          * put_page(page);
472          *                              !page_count(page)   [good, discard it]
473          *
474          * [oops, our write_to data is lost]
475          *
476          * Reversing the order of the tests ensures such a situation cannot
477          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
478          * load is not satisfied before that of page->_count.
479          *
480          * Note that if SetPageDirty is always performed via set_page_dirty,
481          * and thus under tree_lock, then this ordering is not required.
482          */
483         if (!page_freeze_refs(page, 2))
484                 goto cannot_free;
485         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
486         if (unlikely(PageDirty(page))) {
487                 page_unfreeze_refs(page, 2);
488                 goto cannot_free;
489         }
490
491         if (PageSwapCache(page)) {
492                 swp_entry_t swap = { .val = page_private(page) };
493                 __delete_from_swap_cache(page);
494                 spin_unlock_irq(&mapping->tree_lock);
495                 swapcache_free(swap, page);
496         } else {
497                 void (*freepage)(struct page *);
498
499                 freepage = mapping->a_ops->freepage;
500
501                 __remove_from_page_cache(page);
502                 spin_unlock_irq(&mapping->tree_lock);
503                 mem_cgroup_uncharge_cache_page(page);
504
505                 if (freepage != NULL)
506                         freepage(page);
507         }
508
509         return 1;
510
511 cannot_free:
512         spin_unlock_irq(&mapping->tree_lock);
513         return 0;
514 }
515
516 /*
517  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
518  * someone else has a ref on the page, abort and return 0.  If it was
519  * successfully detached, return 1.  Assumes the caller has a single ref on
520  * this page.
521  */
522 int remove_mapping(struct address_space *mapping, struct page *page)
523 {
524         if (__remove_mapping(mapping, page)) {
525                 /*
526                  * Unfreezing the refcount with 1 rather than 2 effectively
527                  * drops the pagecache ref for us without requiring another
528                  * atomic operation.
529                  */
530                 page_unfreeze_refs(page, 1);
531                 return 1;
532         }
533         return 0;
534 }
535
536 /**
537  * putback_lru_page - put previously isolated page onto appropriate LRU list
538  * @page: page to be put back to appropriate lru list
539  *
540  * Add previously isolated @page to appropriate LRU list.
541  * Page may still be unevictable for other reasons.
542  *
543  * lru_lock must not be held, interrupts must be enabled.
544  */
545 void putback_lru_page(struct page *page)
546 {
547         int lru;
548         int active = !!TestClearPageActive(page);
549         int was_unevictable = PageUnevictable(page);
550
551         VM_BUG_ON(PageLRU(page));
552
553 redo:
554         ClearPageUnevictable(page);
555
556         if (page_evictable(page, NULL)) {
557                 /*
558                  * For evictable pages, we can use the cache.
559                  * In event of a race, worst case is we end up with an
560                  * unevictable page on [in]active list.
561                  * We know how to handle that.
562                  */
563                 lru = active + page_lru_base_type(page);
564                 lru_cache_add_lru(page, lru);
565         } else {
566                 /*
567                  * Put unevictable pages directly on zone's unevictable
568                  * list.
569                  */
570                 lru = LRU_UNEVICTABLE;
571                 add_page_to_unevictable_list(page);
572                 /*
573                  * When racing with an mlock clearing (page is
574                  * unlocked), make sure that if the other thread does
575                  * not observe our setting of PG_lru and fails
576                  * isolation, we see PG_mlocked cleared below and move
577                  * the page back to the evictable list.
578                  *
579                  * The other side is TestClearPageMlocked().
580                  */
581                 smp_mb();
582         }
583
584         /*
585          * page's status can change while we move it among lru. If an evictable
586          * page is on unevictable list, it never be freed. To avoid that,
587          * check after we added it to the list, again.
588          */
589         if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
590                 if (!isolate_lru_page(page)) {
591                         put_page(page);
592                         goto redo;
593                 }
594                 /* This means someone else dropped this page from LRU
595                  * So, it will be freed or putback to LRU again. There is
596                  * nothing to do here.
597                  */
598         }
599
600         if (was_unevictable && lru != LRU_UNEVICTABLE)
601                 count_vm_event(UNEVICTABLE_PGRESCUED);
602         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
603                 count_vm_event(UNEVICTABLE_PGCULLED);
604
605         put_page(page);         /* drop ref from isolate */
606 }
607
608 enum page_references {
609         PAGEREF_RECLAIM,
610         PAGEREF_RECLAIM_CLEAN,
611         PAGEREF_KEEP,
612         PAGEREF_ACTIVATE,
613 };
614
615 static enum page_references page_check_references(struct page *page,
616                                                   struct scan_control *sc)
617 {
618         int referenced_ptes, referenced_page;
619         unsigned long vm_flags;
620
621         referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
622         referenced_page = TestClearPageReferenced(page);
623
624         /* Lumpy reclaim - ignore references */
625         if (sc->lumpy_reclaim_mode != LUMPY_MODE_NONE)
626                 return PAGEREF_RECLAIM;
627
628         /*
629          * Mlock lost the isolation race with us.  Let try_to_unmap()
630          * move the page to the unevictable list.
631          */
632         if (vm_flags & VM_LOCKED)
633                 return PAGEREF_RECLAIM;
634
635         if (referenced_ptes) {
636                 if (PageAnon(page))
637                         return PAGEREF_ACTIVATE;
638                 /*
639                  * All mapped pages start out with page table
640                  * references from the instantiating fault, so we need
641                  * to look twice if a mapped file page is used more
642                  * than once.
643                  *
644                  * Mark it and spare it for another trip around the
645                  * inactive list.  Another page table reference will
646                  * lead to its activation.
647                  *
648                  * Note: the mark is set for activated pages as well
649                  * so that recently deactivated but used pages are
650                  * quickly recovered.
651                  */
652                 SetPageReferenced(page);
653
654                 if (referenced_page)
655                         return PAGEREF_ACTIVATE;
656
657                 return PAGEREF_KEEP;
658         }
659
660         /* Reclaim if clean, defer dirty pages to writeback */
661         if (referenced_page && !PageSwapBacked(page))
662                 return PAGEREF_RECLAIM_CLEAN;
663
664         return PAGEREF_RECLAIM;
665 }
666
667 static noinline_for_stack void free_page_list(struct list_head *free_pages)
668 {
669         struct pagevec freed_pvec;
670         struct page *page, *tmp;
671
672         pagevec_init(&freed_pvec, 1);
673
674         list_for_each_entry_safe(page, tmp, free_pages, lru) {
675                 list_del(&page->lru);
676                 if (!pagevec_add(&freed_pvec, page)) {
677                         __pagevec_free(&freed_pvec);
678                         pagevec_reinit(&freed_pvec);
679                 }
680         }
681
682         pagevec_free(&freed_pvec);
683 }
684
685 /*
686  * shrink_page_list() returns the number of reclaimed pages
687  */
688 static unsigned long shrink_page_list(struct list_head *page_list,
689                                       struct zone *zone,
690                                       struct scan_control *sc)
691 {
692         LIST_HEAD(ret_pages);
693         LIST_HEAD(free_pages);
694         int pgactivate = 0;
695         unsigned long nr_dirty = 0;
696         unsigned long nr_congested = 0;
697         unsigned long nr_reclaimed = 0;
698
699         cond_resched();
700
701         while (!list_empty(page_list)) {
702                 enum page_references references;
703                 struct address_space *mapping;
704                 struct page *page;
705                 int may_enter_fs;
706
707                 cond_resched();
708
709                 page = lru_to_page(page_list);
710                 list_del(&page->lru);
711
712                 if (!trylock_page(page))
713                         goto keep;
714
715                 VM_BUG_ON(PageActive(page));
716                 VM_BUG_ON(page_zone(page) != zone);
717
718                 sc->nr_scanned++;
719
720                 if (unlikely(!page_evictable(page, NULL)))
721                         goto cull_mlocked;
722
723                 if (!sc->may_unmap && page_mapped(page))
724                         goto keep_locked;
725
726                 /* Double the slab pressure for mapped and swapcache pages */
727                 if (page_mapped(page) || PageSwapCache(page))
728                         sc->nr_scanned++;
729
730                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
731                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
732
733                 if (PageWriteback(page)) {
734                         /*
735                          * Synchronous reclaim is performed in two passes,
736                          * first an asynchronous pass over the list to
737                          * start parallel writeback, and a second synchronous
738                          * pass to wait for the IO to complete.  Wait here
739                          * for any page for which writeback has already
740                          * started.
741                          */
742                         if (sc->lumpy_reclaim_mode == LUMPY_MODE_SYNC &&
743                             may_enter_fs)
744                                 wait_on_page_writeback(page);
745                         else {
746                                 unlock_page(page);
747                                 goto keep_lumpy;
748                         }
749                 }
750
751                 references = page_check_references(page, sc);
752                 switch (references) {
753                 case PAGEREF_ACTIVATE:
754                         goto activate_locked;
755                 case PAGEREF_KEEP:
756                         goto keep_locked;
757                 case PAGEREF_RECLAIM:
758                 case PAGEREF_RECLAIM_CLEAN:
759                         ; /* try to reclaim the page below */
760                 }
761
762                 /*
763                  * Anonymous process memory has backing store?
764                  * Try to allocate it some swap space here.
765                  */
766                 if (PageAnon(page) && !PageSwapCache(page)) {
767                         if (!(sc->gfp_mask & __GFP_IO))
768                                 goto keep_locked;
769                         if (!add_to_swap(page))
770                                 goto activate_locked;
771                         may_enter_fs = 1;
772                 }
773
774                 mapping = page_mapping(page);
775
776                 /*
777                  * The page is mapped into the page tables of one or more
778                  * processes. Try to unmap it here.
779                  */
780                 if (page_mapped(page) && mapping) {
781                         switch (try_to_unmap(page, TTU_UNMAP)) {
782                         case SWAP_FAIL:
783                                 goto activate_locked;
784                         case SWAP_AGAIN:
785                                 goto keep_locked;
786                         case SWAP_MLOCK:
787                                 goto cull_mlocked;
788                         case SWAP_SUCCESS:
789                                 ; /* try to free the page below */
790                         }
791                 }
792
793                 if (PageDirty(page)) {
794                         nr_dirty++;
795
796                         if (references == PAGEREF_RECLAIM_CLEAN)
797                                 goto keep_locked;
798                         if (!may_enter_fs)
799                                 goto keep_locked;
800                         if (!sc->may_writepage)
801                                 goto keep_locked;
802
803                         /* Page is dirty, try to write it out here */
804                         switch (pageout(page, mapping, sc)) {
805                         case PAGE_KEEP:
806                                 nr_congested++;
807                                 goto keep_locked;
808                         case PAGE_ACTIVATE:
809                                 goto activate_locked;
810                         case PAGE_SUCCESS:
811                                 if (PageWriteback(page))
812                                         goto keep_lumpy;
813                                 if (PageDirty(page))
814                                         goto keep;
815
816                                 /*
817                                  * A synchronous write - probably a ramdisk.  Go
818                                  * ahead and try to reclaim the page.
819                                  */
820                                 if (!trylock_page(page))
821                                         goto keep;
822                                 if (PageDirty(page) || PageWriteback(page))
823                                         goto keep_locked;
824                                 mapping = page_mapping(page);
825                         case PAGE_CLEAN:
826                                 ; /* try to free the page below */
827                         }
828                 }
829
830                 /*
831                  * If the page has buffers, try to free the buffer mappings
832                  * associated with this page. If we succeed we try to free
833                  * the page as well.
834                  *
835                  * We do this even if the page is PageDirty().
836                  * try_to_release_page() does not perform I/O, but it is
837                  * possible for a page to have PageDirty set, but it is actually
838                  * clean (all its buffers are clean).  This happens if the
839                  * buffers were written out directly, with submit_bh(). ext3
840                  * will do this, as well as the blockdev mapping.
841                  * try_to_release_page() will discover that cleanness and will
842                  * drop the buffers and mark the page clean - it can be freed.
843                  *
844                  * Rarely, pages can have buffers and no ->mapping.  These are
845                  * the pages which were not successfully invalidated in
846                  * truncate_complete_page().  We try to drop those buffers here
847                  * and if that worked, and the page is no longer mapped into
848                  * process address space (page_count == 1) it can be freed.
849                  * Otherwise, leave the page on the LRU so it is swappable.
850                  */
851                 if (page_has_private(page)) {
852                         if (!try_to_release_page(page, sc->gfp_mask))
853                                 goto activate_locked;
854                         if (!mapping && page_count(page) == 1) {
855                                 unlock_page(page);
856                                 if (put_page_testzero(page))
857                                         goto free_it;
858                                 else {
859                                         /*
860                                          * rare race with speculative reference.
861                                          * the speculative reference will free
862                                          * this page shortly, so we may
863                                          * increment nr_reclaimed here (and
864                                          * leave it off the LRU).
865                                          */
866                                         nr_reclaimed++;
867                                         continue;
868                                 }
869                         }
870                 }
871
872                 if (!mapping || !__remove_mapping(mapping, page))
873                         goto keep_locked;
874
875                 /*
876                  * At this point, we have no other references and there is
877                  * no way to pick any more up (removed from LRU, removed
878                  * from pagecache). Can use non-atomic bitops now (and
879                  * we obviously don't have to worry about waking up a process
880                  * waiting on the page lock, because there are no references.
881                  */
882                 __clear_page_locked(page);
883 free_it:
884                 nr_reclaimed++;
885
886                 /*
887                  * Is there need to periodically free_page_list? It would
888                  * appear not as the counts should be low
889                  */
890                 list_add(&page->lru, &free_pages);
891                 continue;
892
893 cull_mlocked:
894                 if (PageSwapCache(page))
895                         try_to_free_swap(page);
896                 unlock_page(page);
897                 putback_lru_page(page);
898                 disable_lumpy_reclaim_mode(sc);
899                 continue;
900
901 activate_locked:
902                 /* Not a candidate for swapping, so reclaim swap space. */
903                 if (PageSwapCache(page) && vm_swap_full())
904                         try_to_free_swap(page);
905                 VM_BUG_ON(PageActive(page));
906                 SetPageActive(page);
907                 pgactivate++;
908 keep_locked:
909                 unlock_page(page);
910 keep:
911                 disable_lumpy_reclaim_mode(sc);
912 keep_lumpy:
913                 list_add(&page->lru, &ret_pages);
914                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
915         }
916
917         /*
918          * Tag a zone as congested if all the dirty pages encountered were
919          * backed by a congested BDI. In this case, reclaimers should just
920          * back off and wait for congestion to clear because further reclaim
921          * will encounter the same problem
922          */
923         if (nr_dirty == nr_congested && nr_dirty != 0)
924                 zone_set_flag(zone, ZONE_CONGESTED);
925
926         free_page_list(&free_pages);
927
928         list_splice(&ret_pages, page_list);
929         count_vm_events(PGACTIVATE, pgactivate);
930         return nr_reclaimed;
931 }
932
933 /*
934  * Attempt to remove the specified page from its LRU.  Only take this page
935  * if it is of the appropriate PageActive status.  Pages which are being
936  * freed elsewhere are also ignored.
937  *
938  * page:        page to consider
939  * mode:        one of the LRU isolation modes defined above
940  *
941  * returns 0 on success, -ve errno on failure.
942  */
943 int __isolate_lru_page(struct page *page, int mode, int file)
944 {
945         int ret = -EINVAL;
946
947         /* Only take pages on the LRU. */
948         if (!PageLRU(page))
949                 return ret;
950
951         /*
952          * When checking the active state, we need to be sure we are
953          * dealing with comparible boolean values.  Take the logical not
954          * of each.
955          */
956         if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
957                 return ret;
958
959         if (mode != ISOLATE_BOTH && page_is_file_cache(page) != file)
960                 return ret;
961
962         /*
963          * When this function is being called for lumpy reclaim, we
964          * initially look into all LRU pages, active, inactive and
965          * unevictable; only give shrink_page_list evictable pages.
966          */
967         if (PageUnevictable(page))
968                 return ret;
969
970         ret = -EBUSY;
971
972         if (likely(get_page_unless_zero(page))) {
973                 /*
974                  * Be careful not to clear PageLRU until after we're
975                  * sure the page is not being freed elsewhere -- the
976                  * page release code relies on it.
977                  */
978                 ClearPageLRU(page);
979                 ret = 0;
980         }
981
982         return ret;
983 }
984
985 /*
986  * zone->lru_lock is heavily contended.  Some of the functions that
987  * shrink the lists perform better by taking out a batch of pages
988  * and working on them outside the LRU lock.
989  *
990  * For pagecache intensive workloads, this function is the hottest
991  * spot in the kernel (apart from copy_*_user functions).
992  *
993  * Appropriate locks must be held before calling this function.
994  *
995  * @nr_to_scan: The number of pages to look through on the list.
996  * @src:        The LRU list to pull pages off.
997  * @dst:        The temp list to put pages on to.
998  * @scanned:    The number of pages that were scanned.
999  * @order:      The caller's attempted allocation order
1000  * @mode:       One of the LRU isolation modes
1001  * @file:       True [1] if isolating file [!anon] pages
1002  *
1003  * returns how many pages were moved onto *@dst.
1004  */
1005 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1006                 struct list_head *src, struct list_head *dst,
1007                 unsigned long *scanned, int order, int mode, int file)
1008 {
1009         unsigned long nr_taken = 0;
1010         unsigned long nr_lumpy_taken = 0;
1011         unsigned long nr_lumpy_dirty = 0;
1012         unsigned long nr_lumpy_failed = 0;
1013         unsigned long scan;
1014
1015         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1016                 struct page *page;
1017                 unsigned long pfn;
1018                 unsigned long end_pfn;
1019                 unsigned long page_pfn;
1020                 int zone_id;
1021
1022                 page = lru_to_page(src);
1023                 prefetchw_prev_lru_page(page, src, flags);
1024
1025                 VM_BUG_ON(!PageLRU(page));
1026
1027                 switch (__isolate_lru_page(page, mode, file)) {
1028                 case 0:
1029                         list_move(&page->lru, dst);
1030                         mem_cgroup_del_lru(page);
1031                         nr_taken++;
1032                         break;
1033
1034                 case -EBUSY:
1035                         /* else it is being freed elsewhere */
1036                         list_move(&page->lru, src);
1037                         mem_cgroup_rotate_lru_list(page, page_lru(page));
1038                         continue;
1039
1040                 default:
1041                         BUG();
1042                 }
1043
1044                 if (!order)
1045                         continue;
1046
1047                 /*
1048                  * Attempt to take all pages in the order aligned region
1049                  * surrounding the tag page.  Only take those pages of
1050                  * the same active state as that tag page.  We may safely
1051                  * round the target page pfn down to the requested order
1052                  * as the mem_map is guarenteed valid out to MAX_ORDER,
1053                  * where that page is in a different zone we will detect
1054                  * it from its zone id and abort this block scan.
1055                  */
1056                 zone_id = page_zone_id(page);
1057                 page_pfn = page_to_pfn(page);
1058                 pfn = page_pfn & ~((1 << order) - 1);
1059                 end_pfn = pfn + (1 << order);
1060                 for (; pfn < end_pfn; pfn++) {
1061                         struct page *cursor_page;
1062
1063                         /* The target page is in the block, ignore it. */
1064                         if (unlikely(pfn == page_pfn))
1065                                 continue;
1066
1067                         /* Avoid holes within the zone. */
1068                         if (unlikely(!pfn_valid_within(pfn)))
1069                                 break;
1070
1071                         cursor_page = pfn_to_page(pfn);
1072
1073                         /* Check that we have not crossed a zone boundary. */
1074                         if (unlikely(page_zone_id(cursor_page) != zone_id))
1075                                 break;
1076
1077                         /*
1078                          * If we don't have enough swap space, reclaiming of
1079                          * anon page which don't already have a swap slot is
1080                          * pointless.
1081                          */
1082                         if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
1083                             !PageSwapCache(cursor_page))
1084                                 break;
1085
1086                         if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1087                                 list_move(&cursor_page->lru, dst);
1088                                 mem_cgroup_del_lru(cursor_page);
1089                                 nr_taken++;
1090                                 nr_lumpy_taken++;
1091                                 if (PageDirty(cursor_page))
1092                                         nr_lumpy_dirty++;
1093                                 scan++;
1094                         } else {
1095                                 /* the page is freed already. */
1096                                 if (!page_count(cursor_page))
1097                                         continue;
1098                                 break;
1099                         }
1100                 }
1101
1102                 /* If we break out of the loop above, lumpy reclaim failed */
1103                 if (pfn < end_pfn)
1104                         nr_lumpy_failed++;
1105         }
1106
1107         *scanned = scan;
1108
1109         trace_mm_vmscan_lru_isolate(order,
1110                         nr_to_scan, scan,
1111                         nr_taken,
1112                         nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1113                         mode);
1114         return nr_taken;
1115 }
1116
1117 static unsigned long isolate_pages_global(unsigned long nr,
1118                                         struct list_head *dst,
1119                                         unsigned long *scanned, int order,
1120                                         int mode, struct zone *z,
1121                                         int active, int file)
1122 {
1123         int lru = LRU_BASE;
1124         if (active)
1125                 lru += LRU_ACTIVE;
1126         if (file)
1127                 lru += LRU_FILE;
1128         return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1129                                                                 mode, file);
1130 }
1131
1132 /*
1133  * clear_active_flags() is a helper for shrink_active_list(), clearing
1134  * any active bits from the pages in the list.
1135  */
1136 static unsigned long clear_active_flags(struct list_head *page_list,
1137                                         unsigned int *count)
1138 {
1139         int nr_active = 0;
1140         int lru;
1141         struct page *page;
1142
1143         list_for_each_entry(page, page_list, lru) {
1144                 lru = page_lru_base_type(page);
1145                 if (PageActive(page)) {
1146                         lru += LRU_ACTIVE;
1147                         ClearPageActive(page);
1148                         nr_active++;
1149                 }
1150                 if (count)
1151                         count[lru]++;
1152         }
1153
1154         return nr_active;
1155 }
1156
1157 /**
1158  * isolate_lru_page - tries to isolate a page from its LRU list
1159  * @page: page to isolate from its LRU list
1160  *
1161  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1162  * vmstat statistic corresponding to whatever LRU list the page was on.
1163  *
1164  * Returns 0 if the page was removed from an LRU list.
1165  * Returns -EBUSY if the page was not on an LRU list.
1166  *
1167  * The returned page will have PageLRU() cleared.  If it was found on
1168  * the active list, it will have PageActive set.  If it was found on
1169  * the unevictable list, it will have the PageUnevictable bit set. That flag
1170  * may need to be cleared by the caller before letting the page go.
1171  *
1172  * The vmstat statistic corresponding to the list on which the page was
1173  * found will be decremented.
1174  *
1175  * Restrictions:
1176  * (1) Must be called with an elevated refcount on the page. This is a
1177  *     fundamentnal difference from isolate_lru_pages (which is called
1178  *     without a stable reference).
1179  * (2) the lru_lock must not be held.
1180  * (3) interrupts must be enabled.
1181  */
1182 int isolate_lru_page(struct page *page)
1183 {
1184         int ret = -EBUSY;
1185
1186         if (PageLRU(page)) {
1187                 struct zone *zone = page_zone(page);
1188
1189                 spin_lock_irq(&zone->lru_lock);
1190                 if (PageLRU(page) && get_page_unless_zero(page)) {
1191                         int lru = page_lru(page);
1192                         ret = 0;
1193                         ClearPageLRU(page);
1194
1195                         del_page_from_lru_list(zone, page, lru);
1196                 }
1197                 spin_unlock_irq(&zone->lru_lock);
1198         }
1199         return ret;
1200 }
1201
1202 /*
1203  * Are there way too many processes in the direct reclaim path already?
1204  */
1205 static int too_many_isolated(struct zone *zone, int file,
1206                 struct scan_control *sc)
1207 {
1208         unsigned long inactive, isolated;
1209
1210         if (current_is_kswapd())
1211                 return 0;
1212
1213         if (!scanning_global_lru(sc))
1214                 return 0;
1215
1216         if (file) {
1217                 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1218                 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1219         } else {
1220                 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1221                 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1222         }
1223
1224         return isolated > inactive;
1225 }
1226
1227 /*
1228  * TODO: Try merging with migrations version of putback_lru_pages
1229  */
1230 static noinline_for_stack void
1231 putback_lru_pages(struct zone *zone, struct scan_control *sc,
1232                                 unsigned long nr_anon, unsigned long nr_file,
1233                                 struct list_head *page_list)
1234 {
1235         struct page *page;
1236         struct pagevec pvec;
1237         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1238
1239         pagevec_init(&pvec, 1);
1240
1241         /*
1242          * Put back any unfreeable pages.
1243          */
1244         spin_lock(&zone->lru_lock);
1245         while (!list_empty(page_list)) {
1246                 int lru;
1247                 page = lru_to_page(page_list);
1248                 VM_BUG_ON(PageLRU(page));
1249                 list_del(&page->lru);
1250                 if (unlikely(!page_evictable(page, NULL))) {
1251                         spin_unlock_irq(&zone->lru_lock);
1252                         putback_lru_page(page);
1253                         spin_lock_irq(&zone->lru_lock);
1254                         continue;
1255                 }
1256                 SetPageLRU(page);
1257                 lru = page_lru(page);
1258                 add_page_to_lru_list(zone, page, lru);
1259                 if (is_active_lru(lru)) {
1260                         int file = is_file_lru(lru);
1261                         reclaim_stat->recent_rotated[file]++;
1262                 }
1263                 if (!pagevec_add(&pvec, page)) {
1264                         spin_unlock_irq(&zone->lru_lock);
1265                         __pagevec_release(&pvec);
1266                         spin_lock_irq(&zone->lru_lock);
1267                 }
1268         }
1269         __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1270         __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1271
1272         spin_unlock_irq(&zone->lru_lock);
1273         pagevec_release(&pvec);
1274 }
1275
1276 static noinline_for_stack void update_isolated_counts(struct zone *zone,
1277                                         struct scan_control *sc,
1278                                         unsigned long *nr_anon,
1279                                         unsigned long *nr_file,
1280                                         struct list_head *isolated_list)
1281 {
1282         unsigned long nr_active;
1283         unsigned int count[NR_LRU_LISTS] = { 0, };
1284         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1285
1286         nr_active = clear_active_flags(isolated_list, count);
1287         __count_vm_events(PGDEACTIVATE, nr_active);
1288
1289         __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1290                               -count[LRU_ACTIVE_FILE]);
1291         __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1292                               -count[LRU_INACTIVE_FILE]);
1293         __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1294                               -count[LRU_ACTIVE_ANON]);
1295         __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1296                               -count[LRU_INACTIVE_ANON]);
1297
1298         *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1299         *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1300         __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1301         __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1302
1303         reclaim_stat->recent_scanned[0] += *nr_anon;
1304         reclaim_stat->recent_scanned[1] += *nr_file;
1305 }
1306
1307 /*
1308  * Returns true if the caller should wait to clean dirty/writeback pages.
1309  *
1310  * If we are direct reclaiming for contiguous pages and we do not reclaim
1311  * everything in the list, try again and wait for writeback IO to complete.
1312  * This will stall high-order allocations noticeably. Only do that when really
1313  * need to free the pages under high memory pressure.
1314  */
1315 static inline bool should_reclaim_stall(unsigned long nr_taken,
1316                                         unsigned long nr_freed,
1317                                         int priority,
1318                                         struct scan_control *sc)
1319 {
1320         int lumpy_stall_priority;
1321
1322         /* kswapd should not stall on sync IO */
1323         if (current_is_kswapd())
1324                 return false;
1325
1326         /* Only stall on lumpy reclaim */
1327         if (sc->lumpy_reclaim_mode == LUMPY_MODE_NONE)
1328                 return false;
1329
1330         /* If we have relaimed everything on the isolated list, no stall */
1331         if (nr_freed == nr_taken)
1332                 return false;
1333
1334         /*
1335          * For high-order allocations, there are two stall thresholds.
1336          * High-cost allocations stall immediately where as lower
1337          * order allocations such as stacks require the scanning
1338          * priority to be much higher before stalling.
1339          */
1340         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1341                 lumpy_stall_priority = DEF_PRIORITY;
1342         else
1343                 lumpy_stall_priority = DEF_PRIORITY / 3;
1344
1345         return priority <= lumpy_stall_priority;
1346 }
1347
1348 /*
1349  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1350  * of reclaimed pages
1351  */
1352 static noinline_for_stack unsigned long
1353 shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1354                         struct scan_control *sc, int priority, int file)
1355 {
1356         LIST_HEAD(page_list);
1357         unsigned long nr_scanned;
1358         unsigned long nr_reclaimed = 0;
1359         unsigned long nr_taken;
1360         unsigned long nr_anon;
1361         unsigned long nr_file;
1362
1363         while (unlikely(too_many_isolated(zone, file, sc))) {
1364                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1365
1366                 /* We are about to die and free our memory. Return now. */
1367                 if (fatal_signal_pending(current))
1368                         return SWAP_CLUSTER_MAX;
1369         }
1370
1371         set_lumpy_reclaim_mode(priority, sc, false);
1372         lru_add_drain();
1373         spin_lock_irq(&zone->lru_lock);
1374
1375         if (scanning_global_lru(sc)) {
1376                 nr_taken = isolate_pages_global(nr_to_scan,
1377                         &page_list, &nr_scanned, sc->order,
1378                         sc->lumpy_reclaim_mode == LUMPY_MODE_NONE ?
1379                                         ISOLATE_INACTIVE : ISOLATE_BOTH,
1380                         zone, 0, file);
1381                 zone->pages_scanned += nr_scanned;
1382                 if (current_is_kswapd())
1383                         __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1384                                                nr_scanned);
1385                 else
1386                         __count_zone_vm_events(PGSCAN_DIRECT, zone,
1387                                                nr_scanned);
1388         } else {
1389                 nr_taken = mem_cgroup_isolate_pages(nr_to_scan,
1390                         &page_list, &nr_scanned, sc->order,
1391                         sc->lumpy_reclaim_mode == LUMPY_MODE_NONE ?
1392                                         ISOLATE_INACTIVE : ISOLATE_BOTH,
1393                         zone, sc->mem_cgroup,
1394                         0, file);
1395                 /*
1396                  * mem_cgroup_isolate_pages() keeps track of
1397                  * scanned pages on its own.
1398                  */
1399         }
1400
1401         if (nr_taken == 0) {
1402                 spin_unlock_irq(&zone->lru_lock);
1403                 return 0;
1404         }
1405
1406         update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
1407
1408         spin_unlock_irq(&zone->lru_lock);
1409
1410         nr_reclaimed = shrink_page_list(&page_list, zone, sc);
1411
1412         /* Check if we should syncronously wait for writeback */
1413         if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1414                 set_lumpy_reclaim_mode(priority, sc, true);
1415                 nr_reclaimed += shrink_page_list(&page_list, zone, sc);
1416         }
1417
1418         local_irq_disable();
1419         if (current_is_kswapd())
1420                 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1421         __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1422
1423         putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
1424
1425         trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1426                 zone_idx(zone),
1427                 nr_scanned, nr_reclaimed,
1428                 priority,
1429                 trace_shrink_flags(file, sc->lumpy_reclaim_mode));
1430         return nr_reclaimed;
1431 }
1432
1433 /*
1434  * This moves pages from the active list to the inactive list.
1435  *
1436  * We move them the other way if the page is referenced by one or more
1437  * processes, from rmap.
1438  *
1439  * If the pages are mostly unmapped, the processing is fast and it is
1440  * appropriate to hold zone->lru_lock across the whole operation.  But if
1441  * the pages are mapped, the processing is slow (page_referenced()) so we
1442  * should drop zone->lru_lock around each page.  It's impossible to balance
1443  * this, so instead we remove the pages from the LRU while processing them.
1444  * It is safe to rely on PG_active against the non-LRU pages in here because
1445  * nobody will play with that bit on a non-LRU page.
1446  *
1447  * The downside is that we have to touch page->_count against each page.
1448  * But we had to alter page->flags anyway.
1449  */
1450
1451 static void move_active_pages_to_lru(struct zone *zone,
1452                                      struct list_head *list,
1453                                      enum lru_list lru)
1454 {
1455         unsigned long pgmoved = 0;
1456         struct pagevec pvec;
1457         struct page *page;
1458
1459         pagevec_init(&pvec, 1);
1460
1461         while (!list_empty(list)) {
1462                 page = lru_to_page(list);
1463
1464                 VM_BUG_ON(PageLRU(page));
1465                 SetPageLRU(page);
1466
1467                 list_move(&page->lru, &zone->lru[lru].list);
1468                 mem_cgroup_add_lru_list(page, lru);
1469                 pgmoved++;
1470
1471                 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1472                         spin_unlock_irq(&zone->lru_lock);
1473                         if (buffer_heads_over_limit)
1474                                 pagevec_strip(&pvec);
1475                         __pagevec_release(&pvec);
1476                         spin_lock_irq(&zone->lru_lock);
1477                 }
1478         }
1479         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1480         if (!is_active_lru(lru))
1481                 __count_vm_events(PGDEACTIVATE, pgmoved);
1482 }
1483
1484 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1485                         struct scan_control *sc, int priority, int file)
1486 {
1487         unsigned long nr_taken;
1488         unsigned long pgscanned;
1489         unsigned long vm_flags;
1490         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1491         LIST_HEAD(l_active);
1492         LIST_HEAD(l_inactive);
1493         struct page *page;
1494         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1495         unsigned long nr_rotated = 0;
1496
1497         lru_add_drain();
1498         spin_lock_irq(&zone->lru_lock);
1499         if (scanning_global_lru(sc)) {
1500                 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1501                                                 &pgscanned, sc->order,
1502                                                 ISOLATE_ACTIVE, zone,
1503                                                 1, file);
1504                 zone->pages_scanned += pgscanned;
1505         } else {
1506                 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1507                                                 &pgscanned, sc->order,
1508                                                 ISOLATE_ACTIVE, zone,
1509                                                 sc->mem_cgroup, 1, file);
1510                 /*
1511                  * mem_cgroup_isolate_pages() keeps track of
1512                  * scanned pages on its own.
1513                  */
1514         }
1515
1516         reclaim_stat->recent_scanned[file] += nr_taken;
1517
1518         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1519         if (file)
1520                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1521         else
1522                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1523         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1524         spin_unlock_irq(&zone->lru_lock);
1525
1526         while (!list_empty(&l_hold)) {
1527                 cond_resched();
1528                 page = lru_to_page(&l_hold);
1529                 list_del(&page->lru);
1530
1531                 if (unlikely(!page_evictable(page, NULL))) {
1532                         putback_lru_page(page);
1533                         continue;
1534                 }
1535
1536                 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1537                         nr_rotated++;
1538                         /*
1539                          * Identify referenced, file-backed active pages and
1540                          * give them one more trip around the active list. So
1541                          * that executable code get better chances to stay in
1542                          * memory under moderate memory pressure.  Anon pages
1543                          * are not likely to be evicted by use-once streaming
1544                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1545                          * so we ignore them here.
1546                          */
1547                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1548                                 list_add(&page->lru, &l_active);
1549                                 continue;
1550                         }
1551                 }
1552
1553                 ClearPageActive(page);  /* we are de-activating */
1554                 list_add(&page->lru, &l_inactive);
1555         }
1556
1557         /*
1558          * Move pages back to the lru list.
1559          */
1560         spin_lock_irq(&zone->lru_lock);
1561         /*
1562          * Count referenced pages from currently used mappings as rotated,
1563          * even though only some of them are actually re-activated.  This
1564          * helps balance scan pressure between file and anonymous pages in
1565          * get_scan_ratio.
1566          */
1567         reclaim_stat->recent_rotated[file] += nr_rotated;
1568
1569         move_active_pages_to_lru(zone, &l_active,
1570                                                 LRU_ACTIVE + file * LRU_FILE);
1571         move_active_pages_to_lru(zone, &l_inactive,
1572                                                 LRU_BASE   + file * LRU_FILE);
1573         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1574         spin_unlock_irq(&zone->lru_lock);
1575 }
1576
1577 #ifdef CONFIG_SWAP
1578 static int inactive_anon_is_low_global(struct zone *zone)
1579 {
1580         unsigned long active, inactive;
1581
1582         active = zone_page_state(zone, NR_ACTIVE_ANON);
1583         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1584
1585         if (inactive * zone->inactive_ratio < active)
1586                 return 1;
1587
1588         return 0;
1589 }
1590
1591 /**
1592  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1593  * @zone: zone to check
1594  * @sc:   scan control of this context
1595  *
1596  * Returns true if the zone does not have enough inactive anon pages,
1597  * meaning some active anon pages need to be deactivated.
1598  */
1599 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1600 {
1601         int low;
1602
1603         /*
1604          * If we don't have swap space, anonymous page deactivation
1605          * is pointless.
1606          */
1607         if (!total_swap_pages)
1608                 return 0;
1609
1610         if (scanning_global_lru(sc))
1611                 low = inactive_anon_is_low_global(zone);
1612         else
1613                 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1614         return low;
1615 }
1616 #else
1617 static inline int inactive_anon_is_low(struct zone *zone,
1618                                         struct scan_control *sc)
1619 {
1620         return 0;
1621 }
1622 #endif
1623
1624 static int inactive_file_is_low_global(struct zone *zone)
1625 {
1626         unsigned long active, inactive;
1627
1628         active = zone_page_state(zone, NR_ACTIVE_FILE);
1629         inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1630
1631         return (active > inactive);
1632 }
1633
1634 /**
1635  * inactive_file_is_low - check if file pages need to be deactivated
1636  * @zone: zone to check
1637  * @sc:   scan control of this context
1638  *
1639  * When the system is doing streaming IO, memory pressure here
1640  * ensures that active file pages get deactivated, until more
1641  * than half of the file pages are on the inactive list.
1642  *
1643  * Once we get to that situation, protect the system's working
1644  * set from being evicted by disabling active file page aging.
1645  *
1646  * This uses a different ratio than the anonymous pages, because
1647  * the page cache uses a use-once replacement algorithm.
1648  */
1649 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1650 {
1651         int low;
1652
1653         if (scanning_global_lru(sc))
1654                 low = inactive_file_is_low_global(zone);
1655         else
1656                 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1657         return low;
1658 }
1659
1660 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1661                                 int file)
1662 {
1663         if (file)
1664                 return inactive_file_is_low(zone, sc);
1665         else
1666                 return inactive_anon_is_low(zone, sc);
1667 }
1668
1669 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1670         struct zone *zone, struct scan_control *sc, int priority)
1671 {
1672         int file = is_file_lru(lru);
1673
1674         if (is_active_lru(lru)) {
1675                 if (inactive_list_is_low(zone, sc, file))
1676                     shrink_active_list(nr_to_scan, zone, sc, priority, file);
1677                 return 0;
1678         }
1679
1680         return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1681 }
1682
1683 /*
1684  * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1685  * until we collected @swap_cluster_max pages to scan.
1686  */
1687 static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
1688                                        unsigned long *nr_saved_scan)
1689 {
1690         unsigned long nr;
1691
1692         *nr_saved_scan += nr_to_scan;
1693         nr = *nr_saved_scan;
1694
1695         if (nr >= SWAP_CLUSTER_MAX)
1696                 *nr_saved_scan = 0;
1697         else
1698                 nr = 0;
1699
1700         return nr;
1701 }
1702
1703 /*
1704  * Determine how aggressively the anon and file LRU lists should be
1705  * scanned.  The relative value of each set of LRU lists is determined
1706  * by looking at the fraction of the pages scanned we did rotate back
1707  * onto the active list instead of evict.
1708  *
1709  * nr[0] = anon pages to scan; nr[1] = file pages to scan
1710  */
1711 static void get_scan_count(struct zone *zone, struct scan_control *sc,
1712                                         unsigned long *nr, int priority)
1713 {
1714         unsigned long anon, file, free;
1715         unsigned long anon_prio, file_prio;
1716         unsigned long ap, fp;
1717         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1718         u64 fraction[2], denominator;
1719         enum lru_list l;
1720         int noswap = 0;
1721
1722         /* If we have no swap space, do not bother scanning anon pages. */
1723         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1724                 noswap = 1;
1725                 fraction[0] = 0;
1726                 fraction[1] = 1;
1727                 denominator = 1;
1728                 goto out;
1729         }
1730
1731         anon  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1732                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1733         file  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1734                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1735
1736         if (scanning_global_lru(sc)) {
1737                 free  = zone_page_state(zone, NR_FREE_PAGES);
1738                 /* If we have very few page cache pages,
1739                    force-scan anon pages. */
1740                 if (unlikely(file + free <= high_wmark_pages(zone))) {
1741                         fraction[0] = 1;
1742                         fraction[1] = 0;
1743                         denominator = 1;
1744                         goto out;
1745                 }
1746         }
1747
1748         /*
1749          * With swappiness at 100, anonymous and file have the same priority.
1750          * This scanning priority is essentially the inverse of IO cost.
1751          */
1752         anon_prio = sc->swappiness;
1753         file_prio = 200 - sc->swappiness;
1754
1755         /*
1756          * OK, so we have swap space and a fair amount of page cache
1757          * pages.  We use the recently rotated / recently scanned
1758          * ratios to determine how valuable each cache is.
1759          *
1760          * Because workloads change over time (and to avoid overflow)
1761          * we keep these statistics as a floating average, which ends
1762          * up weighing recent references more than old ones.
1763          *
1764          * anon in [0], file in [1]
1765          */
1766         spin_lock_irq(&zone->lru_lock);
1767         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1768                 reclaim_stat->recent_scanned[0] /= 2;
1769                 reclaim_stat->recent_rotated[0] /= 2;
1770         }
1771
1772         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1773                 reclaim_stat->recent_scanned[1] /= 2;
1774                 reclaim_stat->recent_rotated[1] /= 2;
1775         }
1776
1777         /*
1778          * The amount of pressure on anon vs file pages is inversely
1779          * proportional to the fraction of recently scanned pages on
1780          * each list that were recently referenced and in active use.
1781          */
1782         ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1783         ap /= reclaim_stat->recent_rotated[0] + 1;
1784
1785         fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1786         fp /= reclaim_stat->recent_rotated[1] + 1;
1787         spin_unlock_irq(&zone->lru_lock);
1788
1789         fraction[0] = ap;
1790         fraction[1] = fp;
1791         denominator = ap + fp + 1;
1792 out:
1793         for_each_evictable_lru(l) {
1794                 int file = is_file_lru(l);
1795                 unsigned long scan;
1796
1797                 scan = zone_nr_lru_pages(zone, sc, l);
1798                 if (priority || noswap) {
1799                         scan >>= priority;
1800                         scan = div64_u64(scan * fraction[file], denominator);
1801                 }
1802                 nr[l] = nr_scan_try_batch(scan,
1803                                           &reclaim_stat->nr_saved_scan[l]);
1804         }
1805 }
1806
1807 /*
1808  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1809  */
1810 static void shrink_zone(int priority, struct zone *zone,
1811                                 struct scan_control *sc)
1812 {
1813         unsigned long nr[NR_LRU_LISTS];
1814         unsigned long nr_to_scan;
1815         enum lru_list l;
1816         unsigned long nr_reclaimed = sc->nr_reclaimed;
1817         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1818
1819         get_scan_count(zone, sc, nr, priority);
1820
1821         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1822                                         nr[LRU_INACTIVE_FILE]) {
1823                 for_each_evictable_lru(l) {
1824                         if (nr[l]) {
1825                                 nr_to_scan = min_t(unsigned long,
1826                                                    nr[l], SWAP_CLUSTER_MAX);
1827                                 nr[l] -= nr_to_scan;
1828
1829                                 nr_reclaimed += shrink_list(l, nr_to_scan,
1830                                                             zone, sc, priority);
1831                         }
1832                 }
1833                 /*
1834                  * On large memory systems, scan >> priority can become
1835                  * really large. This is fine for the starting priority;
1836                  * we want to put equal scanning pressure on each zone.
1837                  * However, if the VM has a harder time of freeing pages,
1838                  * with multiple processes reclaiming pages, the total
1839                  * freeing target can get unreasonably large.
1840                  */
1841                 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
1842                         break;
1843         }
1844
1845         sc->nr_reclaimed = nr_reclaimed;
1846
1847         /*
1848          * Even if we did not try to evict anon pages at all, we want to
1849          * rebalance the anon lru active/inactive ratio.
1850          */
1851         if (inactive_anon_is_low(zone, sc))
1852                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1853
1854         throttle_vm_writeout(sc->gfp_mask);
1855 }
1856
1857 /*
1858  * This is the direct reclaim path, for page-allocating processes.  We only
1859  * try to reclaim pages from zones which will satisfy the caller's allocation
1860  * request.
1861  *
1862  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1863  * Because:
1864  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1865  *    allocation or
1866  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1867  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1868  *    zone defense algorithm.
1869  *
1870  * If a zone is deemed to be full of pinned pages then just give it a light
1871  * scan then give up on it.
1872  */
1873 static void shrink_zones(int priority, struct zonelist *zonelist,
1874                                         struct scan_control *sc)
1875 {
1876         struct zoneref *z;
1877         struct zone *zone;
1878
1879         for_each_zone_zonelist_nodemask(zone, z, zonelist,
1880                                         gfp_zone(sc->gfp_mask), sc->nodemask) {
1881                 if (!populated_zone(zone))
1882                         continue;
1883                 /*
1884                  * Take care memory controller reclaiming has small influence
1885                  * to global LRU.
1886                  */
1887                 if (scanning_global_lru(sc)) {
1888                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1889                                 continue;
1890                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1891                                 continue;       /* Let kswapd poll it */
1892                 }
1893
1894                 shrink_zone(priority, zone, sc);
1895         }
1896 }
1897
1898 static bool zone_reclaimable(struct zone *zone)
1899 {
1900         return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
1901 }
1902
1903 /*
1904  * As hibernation is going on, kswapd is freezed so that it can't mark
1905  * the zone into all_unreclaimable. It can't handle OOM during hibernation.
1906  * So let's check zone's unreclaimable in direct reclaim as well as kswapd.
1907  */
1908 static bool all_unreclaimable(struct zonelist *zonelist,
1909                 struct scan_control *sc)
1910 {
1911         struct zoneref *z;
1912         struct zone *zone;
1913         bool all_unreclaimable = true;
1914
1915         for_each_zone_zonelist_nodemask(zone, z, zonelist,
1916                         gfp_zone(sc->gfp_mask), sc->nodemask) {
1917                 if (!populated_zone(zone))
1918                         continue;
1919                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1920                         continue;
1921                 if (zone_reclaimable(zone)) {
1922                         all_unreclaimable = false;
1923                         break;
1924                 }
1925         }
1926
1927         return all_unreclaimable;
1928 }
1929
1930 /*
1931  * This is the main entry point to direct page reclaim.
1932  *
1933  * If a full scan of the inactive list fails to free enough memory then we
1934  * are "out of memory" and something needs to be killed.
1935  *
1936  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1937  * high - the zone may be full of dirty or under-writeback pages, which this
1938  * caller can't do much about.  We kick the writeback threads and take explicit
1939  * naps in the hope that some of these pages can be written.  But if the
1940  * allocating task holds filesystem locks which prevent writeout this might not
1941  * work, and the allocation attempt will fail.
1942  *
1943  * returns:     0, if no pages reclaimed
1944  *              else, the number of pages reclaimed
1945  */
1946 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
1947                                         struct scan_control *sc)
1948 {
1949         int priority;
1950         unsigned long total_scanned = 0;
1951         struct reclaim_state *reclaim_state = current->reclaim_state;
1952         struct zoneref *z;
1953         struct zone *zone;
1954         unsigned long writeback_threshold;
1955
1956         get_mems_allowed();
1957         delayacct_freepages_start();
1958
1959         if (scanning_global_lru(sc))
1960                 count_vm_event(ALLOCSTALL);
1961
1962         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1963                 sc->nr_scanned = 0;
1964                 if (!priority)
1965                         disable_swap_token();
1966                 shrink_zones(priority, zonelist, sc);
1967                 /*
1968                  * Don't shrink slabs when reclaiming memory from
1969                  * over limit cgroups
1970                  */
1971                 if (scanning_global_lru(sc)) {
1972                         unsigned long lru_pages = 0;
1973                         for_each_zone_zonelist(zone, z, zonelist,
1974                                         gfp_zone(sc->gfp_mask)) {
1975                                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1976                                         continue;
1977
1978                                 lru_pages += zone_reclaimable_pages(zone);
1979                         }
1980
1981                         shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
1982                         if (reclaim_state) {
1983                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
1984                                 reclaim_state->reclaimed_slab = 0;
1985                         }
1986                 }
1987                 total_scanned += sc->nr_scanned;
1988                 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
1989                         goto out;
1990
1991                 /*
1992                  * Try to write back as many pages as we just scanned.  This
1993                  * tends to cause slow streaming writers to write data to the
1994                  * disk smoothly, at the dirtying rate, which is nice.   But
1995                  * that's undesirable in laptop mode, where we *want* lumpy
1996                  * writeout.  So in laptop mode, write out the whole world.
1997                  */
1998                 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
1999                 if (total_scanned > writeback_threshold) {
2000                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
2001                         sc->may_writepage = 1;
2002                 }
2003
2004                 /* Take a nap, wait for some writeback to complete */
2005                 if (!sc->hibernation_mode && sc->nr_scanned &&
2006                     priority < DEF_PRIORITY - 2) {
2007                         struct zone *preferred_zone;
2008
2009                         first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2010                                                         NULL, &preferred_zone);
2011                         wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2012                 }
2013         }
2014
2015 out:
2016         delayacct_freepages_end();
2017         put_mems_allowed();
2018
2019         if (sc->nr_reclaimed)
2020                 return sc->nr_reclaimed;
2021
2022         /* top priority shrink_zones still had more to do? don't OOM, then */
2023         if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
2024                 return 1;
2025
2026         return 0;
2027 }
2028
2029 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2030                                 gfp_t gfp_mask, nodemask_t *nodemask)
2031 {
2032         unsigned long nr_reclaimed;
2033         struct scan_control sc = {
2034                 .gfp_mask = gfp_mask,
2035                 .may_writepage = !laptop_mode,
2036                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2037                 .may_unmap = 1,
2038                 .may_swap = 1,
2039                 .swappiness = vm_swappiness,
2040                 .order = order,
2041                 .mem_cgroup = NULL,
2042                 .nodemask = nodemask,
2043         };
2044
2045         trace_mm_vmscan_direct_reclaim_begin(order,
2046                                 sc.may_writepage,
2047                                 gfp_mask);
2048
2049         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2050
2051         trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2052
2053         return nr_reclaimed;
2054 }
2055
2056 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2057
2058 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2059                                                 gfp_t gfp_mask, bool noswap,
2060                                                 unsigned int swappiness,
2061                                                 struct zone *zone)
2062 {
2063         struct scan_control sc = {
2064                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2065                 .may_writepage = !laptop_mode,
2066                 .may_unmap = 1,
2067                 .may_swap = !noswap,
2068                 .swappiness = swappiness,
2069                 .order = 0,
2070                 .mem_cgroup = mem,
2071         };
2072         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2073                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2074
2075         trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2076                                                       sc.may_writepage,
2077                                                       sc.gfp_mask);
2078
2079         /*
2080          * NOTE: Although we can get the priority field, using it
2081          * here is not a good idea, since it limits the pages we can scan.
2082          * if we don't reclaim here, the shrink_zone from balance_pgdat
2083          * will pick up pages from other mem cgroup's as well. We hack
2084          * the priority and make it zero.
2085          */
2086         shrink_zone(0, zone, &sc);
2087
2088         trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2089
2090         return sc.nr_reclaimed;
2091 }
2092
2093 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
2094                                            gfp_t gfp_mask,
2095                                            bool noswap,
2096                                            unsigned int swappiness)
2097 {
2098         struct zonelist *zonelist;
2099         unsigned long nr_reclaimed;
2100         struct scan_control sc = {
2101                 .may_writepage = !laptop_mode,
2102                 .may_unmap = 1,
2103                 .may_swap = !noswap,
2104                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2105                 .swappiness = swappiness,
2106                 .order = 0,
2107                 .mem_cgroup = mem_cont,
2108                 .nodemask = NULL, /* we don't care the placement */
2109         };
2110
2111         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2112                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2113         zonelist = NODE_DATA(numa_node_id())->node_zonelists;
2114
2115         trace_mm_vmscan_memcg_reclaim_begin(0,
2116                                             sc.may_writepage,
2117                                             sc.gfp_mask);
2118
2119         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2120
2121         trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2122
2123         return nr_reclaimed;
2124 }
2125 #endif
2126
2127 /* is kswapd sleeping prematurely? */
2128 static int sleeping_prematurely(pg_data_t *pgdat, int order, long remaining)
2129 {
2130         int i;
2131
2132         /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2133         if (remaining)
2134                 return 1;
2135
2136         /* If after HZ/10, a zone is below the high mark, it's premature */
2137         for (i = 0; i < pgdat->nr_zones; i++) {
2138                 struct zone *zone = pgdat->node_zones + i;
2139
2140                 if (!populated_zone(zone))
2141                         continue;
2142
2143                 if (zone->all_unreclaimable)
2144                         continue;
2145
2146                 if (!zone_watermark_ok(zone, order, high_wmark_pages(zone),
2147                                                                 0, 0))
2148                         return 1;
2149         }
2150
2151         return 0;
2152 }
2153
2154 /*
2155  * For kswapd, balance_pgdat() will work across all this node's zones until
2156  * they are all at high_wmark_pages(zone).
2157  *
2158  * Returns the number of pages which were actually freed.
2159  *
2160  * There is special handling here for zones which are full of pinned pages.
2161  * This can happen if the pages are all mlocked, or if they are all used by
2162  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
2163  * What we do is to detect the case where all pages in the zone have been
2164  * scanned twice and there has been zero successful reclaim.  Mark the zone as
2165  * dead and from now on, only perform a short scan.  Basically we're polling
2166  * the zone for when the problem goes away.
2167  *
2168  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
2169  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2170  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2171  * lower zones regardless of the number of free pages in the lower zones. This
2172  * interoperates with the page allocator fallback scheme to ensure that aging
2173  * of pages is balanced across the zones.
2174  */
2175 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
2176 {
2177         int all_zones_ok;
2178         int priority;
2179         int i;
2180         unsigned long total_scanned;
2181         struct reclaim_state *reclaim_state = current->reclaim_state;
2182         struct scan_control sc = {
2183                 .gfp_mask = GFP_KERNEL,
2184                 .may_unmap = 1,
2185                 .may_swap = 1,
2186                 /*
2187                  * kswapd doesn't want to be bailed out while reclaim. because
2188                  * we want to put equal scanning pressure on each zone.
2189                  */
2190                 .nr_to_reclaim = ULONG_MAX,
2191                 .swappiness = vm_swappiness,
2192                 .order = order,
2193                 .mem_cgroup = NULL,
2194         };
2195 loop_again:
2196         total_scanned = 0;
2197         sc.nr_reclaimed = 0;
2198         sc.may_writepage = !laptop_mode;
2199         count_vm_event(PAGEOUTRUN);
2200
2201         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2202                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
2203                 unsigned long lru_pages = 0;
2204                 int has_under_min_watermark_zone = 0;
2205
2206                 /* The swap token gets in the way of swapout... */
2207                 if (!priority)
2208                         disable_swap_token();
2209
2210                 all_zones_ok = 1;
2211
2212                 /*
2213                  * Scan in the highmem->dma direction for the highest
2214                  * zone which needs scanning
2215                  */
2216                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2217                         struct zone *zone = pgdat->node_zones + i;
2218
2219                         if (!populated_zone(zone))
2220                                 continue;
2221
2222                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2223                                 continue;
2224
2225                         /*
2226                          * Do some background aging of the anon list, to give
2227                          * pages a chance to be referenced before reclaiming.
2228                          */
2229                         if (inactive_anon_is_low(zone, &sc))
2230                                 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2231                                                         &sc, priority, 0);
2232
2233                         if (!zone_watermark_ok(zone, order,
2234                                         high_wmark_pages(zone), 0, 0)) {
2235                                 end_zone = i;
2236                                 break;
2237                         }
2238                 }
2239                 if (i < 0)
2240                         goto out;
2241
2242                 for (i = 0; i <= end_zone; i++) {
2243                         struct zone *zone = pgdat->node_zones + i;
2244
2245                         lru_pages += zone_reclaimable_pages(zone);
2246                 }
2247
2248                 /*
2249                  * Now scan the zone in the dma->highmem direction, stopping
2250                  * at the last zone which needs scanning.
2251                  *
2252                  * We do this because the page allocator works in the opposite
2253                  * direction.  This prevents the page allocator from allocating
2254                  * pages behind kswapd's direction of progress, which would
2255                  * cause too much scanning of the lower zones.
2256                  */
2257                 for (i = 0; i <= end_zone; i++) {
2258                         struct zone *zone = pgdat->node_zones + i;
2259                         int nr_slab;
2260
2261                         if (!populated_zone(zone))
2262                                 continue;
2263
2264                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2265                                 continue;
2266
2267                         sc.nr_scanned = 0;
2268
2269                         /*
2270                          * Call soft limit reclaim before calling shrink_zone.
2271                          * For now we ignore the return value
2272                          */
2273                         mem_cgroup_soft_limit_reclaim(zone, order, sc.gfp_mask);
2274
2275                         /*
2276                          * We put equal pressure on every zone, unless one
2277                          * zone has way too many pages free already.
2278                          */
2279                         if (!zone_watermark_ok(zone, order,
2280                                         8*high_wmark_pages(zone), end_zone, 0))
2281                                 shrink_zone(priority, zone, &sc);
2282                         reclaim_state->reclaimed_slab = 0;
2283                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
2284                                                 lru_pages);
2285                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2286                         total_scanned += sc.nr_scanned;
2287                         if (zone->all_unreclaimable)
2288                                 continue;
2289                         if (nr_slab == 0 && !zone_reclaimable(zone))
2290                                 zone->all_unreclaimable = 1;
2291                         /*
2292                          * If we've done a decent amount of scanning and
2293                          * the reclaim ratio is low, start doing writepage
2294                          * even in laptop mode
2295                          */
2296                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2297                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2298                                 sc.may_writepage = 1;
2299
2300                         if (!zone_watermark_ok(zone, order,
2301                                         high_wmark_pages(zone), end_zone, 0)) {
2302                                 all_zones_ok = 0;
2303                                 /*
2304                                  * We are still under min water mark.  This
2305                                  * means that we have a GFP_ATOMIC allocation
2306                                  * failure risk. Hurry up!
2307                                  */
2308                                 if (!zone_watermark_ok(zone, order,
2309                                             min_wmark_pages(zone), end_zone, 0))
2310                                         has_under_min_watermark_zone = 1;
2311                         } else {
2312                                 /*
2313                                  * If a zone reaches its high watermark,
2314                                  * consider it to be no longer congested. It's
2315                                  * possible there are dirty pages backed by
2316                                  * congested BDIs but as pressure is relieved,
2317                                  * spectulatively avoid congestion waits
2318                                  */
2319                                 zone_clear_flag(zone, ZONE_CONGESTED);
2320                         }
2321
2322                 }
2323                 if (all_zones_ok)
2324                         break;          /* kswapd: all done */
2325                 /*
2326                  * OK, kswapd is getting into trouble.  Take a nap, then take
2327                  * another pass across the zones.
2328                  */
2329                 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2330                         if (has_under_min_watermark_zone)
2331                                 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2332                         else
2333                                 congestion_wait(BLK_RW_ASYNC, HZ/10);
2334                 }
2335
2336                 /*
2337                  * We do this so kswapd doesn't build up large priorities for
2338                  * example when it is freeing in parallel with allocators. It
2339                  * matches the direct reclaim path behaviour in terms of impact
2340                  * on zone->*_priority.
2341                  */
2342                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2343                         break;
2344         }
2345 out:
2346         if (!all_zones_ok) {
2347                 cond_resched();
2348
2349                 try_to_freeze();
2350
2351                 /*
2352                  * Fragmentation may mean that the system cannot be
2353                  * rebalanced for high-order allocations in all zones.
2354                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2355                  * it means the zones have been fully scanned and are still
2356                  * not balanced. For high-order allocations, there is
2357                  * little point trying all over again as kswapd may
2358                  * infinite loop.
2359                  *
2360                  * Instead, recheck all watermarks at order-0 as they
2361                  * are the most important. If watermarks are ok, kswapd will go
2362                  * back to sleep. High-order users can still perform direct
2363                  * reclaim if they wish.
2364                  */
2365                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2366                         order = sc.order = 0;
2367
2368                 goto loop_again;
2369         }
2370
2371         return sc.nr_reclaimed;
2372 }
2373
2374 /*
2375  * The background pageout daemon, started as a kernel thread
2376  * from the init process.
2377  *
2378  * This basically trickles out pages so that we have _some_
2379  * free memory available even if there is no other activity
2380  * that frees anything up. This is needed for things like routing
2381  * etc, where we otherwise might have all activity going on in
2382  * asynchronous contexts that cannot page things out.
2383  *
2384  * If there are applications that are active memory-allocators
2385  * (most normal use), this basically shouldn't matter.
2386  */
2387 static int kswapd(void *p)
2388 {
2389         unsigned long order;
2390         pg_data_t *pgdat = (pg_data_t*)p;
2391         struct task_struct *tsk = current;
2392         DEFINE_WAIT(wait);
2393         struct reclaim_state reclaim_state = {
2394                 .reclaimed_slab = 0,
2395         };
2396         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2397
2398         lockdep_set_current_reclaim_state(GFP_KERNEL);
2399
2400         if (!cpumask_empty(cpumask))
2401                 set_cpus_allowed_ptr(tsk, cpumask);
2402         current->reclaim_state = &reclaim_state;
2403
2404         /*
2405          * Tell the memory management that we're a "memory allocator",
2406          * and that if we need more memory we should get access to it
2407          * regardless (see "__alloc_pages()"). "kswapd" should
2408          * never get caught in the normal page freeing logic.
2409          *
2410          * (Kswapd normally doesn't need memory anyway, but sometimes
2411          * you need a small amount of memory in order to be able to
2412          * page out something else, and this flag essentially protects
2413          * us from recursively trying to free more memory as we're
2414          * trying to free the first piece of memory in the first place).
2415          */
2416         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2417         set_freezable();
2418
2419         order = 0;
2420         for ( ; ; ) {
2421                 unsigned long new_order;
2422                 int ret;
2423
2424                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2425                 new_order = pgdat->kswapd_max_order;
2426                 pgdat->kswapd_max_order = 0;
2427                 if (order < new_order) {
2428                         /*
2429                          * Don't sleep if someone wants a larger 'order'
2430                          * allocation
2431                          */
2432                         order = new_order;
2433                 } else {
2434                         if (!freezing(current) && !kthread_should_stop()) {
2435                                 long remaining = 0;
2436
2437                                 /* Try to sleep for a short interval */
2438                                 if (!sleeping_prematurely(pgdat, order, remaining)) {
2439                                         remaining = schedule_timeout(HZ/10);
2440                                         finish_wait(&pgdat->kswapd_wait, &wait);
2441                                         prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2442                                 }
2443
2444                                 /*
2445                                  * After a short sleep, check if it was a
2446                                  * premature sleep. If not, then go fully
2447                                  * to sleep until explicitly woken up
2448                                  */
2449                                 if (!sleeping_prematurely(pgdat, order, remaining)) {
2450                                         trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2451                                         schedule();
2452                                 } else {
2453                                         if (remaining)
2454                                                 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2455                                         else
2456                                                 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2457                                 }
2458                         }
2459
2460                         order = pgdat->kswapd_max_order;
2461                 }
2462                 finish_wait(&pgdat->kswapd_wait, &wait);
2463
2464                 ret = try_to_freeze();
2465                 if (kthread_should_stop())
2466                         break;
2467
2468                 /*
2469                  * We can speed up thawing tasks if we don't call balance_pgdat
2470                  * after returning from the refrigerator
2471                  */
2472                 if (!ret) {
2473                         trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2474                         balance_pgdat(pgdat, order);
2475                 }
2476         }
2477         return 0;
2478 }
2479
2480 /*
2481  * A zone is low on free memory, so wake its kswapd task to service it.
2482  */
2483 void wakeup_kswapd(struct zone *zone, int order)
2484 {
2485         pg_data_t *pgdat;
2486
2487         if (!populated_zone(zone))
2488                 return;
2489
2490         pgdat = zone->zone_pgdat;
2491         if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
2492                 return;
2493         if (pgdat->kswapd_max_order < order)
2494                 pgdat->kswapd_max_order = order;
2495         trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
2496         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2497                 return;
2498         if (!waitqueue_active(&pgdat->kswapd_wait))
2499                 return;
2500         wake_up_interruptible(&pgdat->kswapd_wait);
2501 }
2502
2503 /*
2504  * The reclaimable count would be mostly accurate.
2505  * The less reclaimable pages may be
2506  * - mlocked pages, which will be moved to unevictable list when encountered
2507  * - mapped pages, which may require several travels to be reclaimed
2508  * - dirty pages, which is not "instantly" reclaimable
2509  */
2510 unsigned long global_reclaimable_pages(void)
2511 {
2512         int nr;
2513
2514         nr = global_page_state(NR_ACTIVE_FILE) +
2515              global_page_state(NR_INACTIVE_FILE);
2516
2517         if (nr_swap_pages > 0)
2518                 nr += global_page_state(NR_ACTIVE_ANON) +
2519                       global_page_state(NR_INACTIVE_ANON);
2520
2521         return nr;
2522 }
2523
2524 unsigned long zone_reclaimable_pages(struct zone *zone)
2525 {
2526         int nr;
2527
2528         nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2529              zone_page_state(zone, NR_INACTIVE_FILE);
2530
2531         if (nr_swap_pages > 0)
2532                 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2533                       zone_page_state(zone, NR_INACTIVE_ANON);
2534
2535         return nr;
2536 }
2537
2538 #ifdef CONFIG_HIBERNATION
2539 /*
2540  * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2541  * freed pages.
2542  *
2543  * Rather than trying to age LRUs the aim is to preserve the overall
2544  * LRU order by reclaiming preferentially
2545  * inactive > active > active referenced > active mapped
2546  */
2547 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
2548 {
2549         struct reclaim_state reclaim_state;
2550         struct scan_control sc = {
2551                 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2552                 .may_swap = 1,
2553                 .may_unmap = 1,
2554                 .may_writepage = 1,
2555                 .nr_to_reclaim = nr_to_reclaim,
2556                 .hibernation_mode = 1,
2557                 .swappiness = vm_swappiness,
2558                 .order = 0,
2559         };
2560         struct zonelist * zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2561         struct task_struct *p = current;
2562         unsigned long nr_reclaimed;
2563
2564         p->flags |= PF_MEMALLOC;
2565         lockdep_set_current_reclaim_state(sc.gfp_mask);
2566         reclaim_state.reclaimed_slab = 0;
2567         p->reclaim_state = &reclaim_state;
2568
2569         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2570
2571         p->reclaim_state = NULL;
2572         lockdep_clear_current_reclaim_state();
2573         p->flags &= ~PF_MEMALLOC;
2574
2575         return nr_reclaimed;
2576 }
2577 #endif /* CONFIG_HIBERNATION */
2578
2579 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2580    not required for correctness.  So if the last cpu in a node goes
2581    away, we get changed to run anywhere: as the first one comes back,
2582    restore their cpu bindings. */
2583 static int __devinit cpu_callback(struct notifier_block *nfb,
2584                                   unsigned long action, void *hcpu)
2585 {
2586         int nid;
2587
2588         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2589                 for_each_node_state(nid, N_HIGH_MEMORY) {
2590                         pg_data_t *pgdat = NODE_DATA(nid);
2591                         const struct cpumask *mask;
2592
2593                         mask = cpumask_of_node(pgdat->node_id);
2594
2595                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2596                                 /* One of our CPUs online: restore mask */
2597                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2598                 }
2599         }
2600         return NOTIFY_OK;
2601 }
2602
2603 /*
2604  * This kswapd start function will be called by init and node-hot-add.
2605  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2606  */
2607 int kswapd_run(int nid)
2608 {
2609         pg_data_t *pgdat = NODE_DATA(nid);
2610         int ret = 0;
2611
2612         if (pgdat->kswapd)
2613                 return 0;
2614
2615         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2616         if (IS_ERR(pgdat->kswapd)) {
2617                 /* failure at boot is fatal */
2618                 BUG_ON(system_state == SYSTEM_BOOTING);
2619                 printk("Failed to start kswapd on node %d\n",nid);
2620                 ret = -1;
2621         }
2622         return ret;
2623 }
2624
2625 /*
2626  * Called by memory hotplug when all memory in a node is offlined.
2627  */
2628 void kswapd_stop(int nid)
2629 {
2630         struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
2631
2632         if (kswapd)
2633                 kthread_stop(kswapd);
2634 }
2635
2636 static int __init kswapd_init(void)
2637 {
2638         int nid;
2639
2640         swap_setup();
2641         for_each_node_state(nid, N_HIGH_MEMORY)
2642                 kswapd_run(nid);
2643         hotcpu_notifier(cpu_callback, 0);
2644         return 0;
2645 }
2646
2647 module_init(kswapd_init)
2648
2649 #ifdef CONFIG_NUMA
2650 /*
2651  * Zone reclaim mode
2652  *
2653  * If non-zero call zone_reclaim when the number of free pages falls below
2654  * the watermarks.
2655  */
2656 int zone_reclaim_mode __read_mostly;
2657
2658 #define RECLAIM_OFF 0
2659 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
2660 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
2661 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
2662
2663 /*
2664  * Priority for ZONE_RECLAIM. This determines the fraction of pages
2665  * of a node considered for each zone_reclaim. 4 scans 1/16th of
2666  * a zone.
2667  */
2668 #define ZONE_RECLAIM_PRIORITY 4
2669
2670 /*
2671  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2672  * occur.
2673  */
2674 int sysctl_min_unmapped_ratio = 1;
2675
2676 /*
2677  * If the number of slab pages in a zone grows beyond this percentage then
2678  * slab reclaim needs to occur.
2679  */
2680 int sysctl_min_slab_ratio = 5;
2681
2682 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
2683 {
2684         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
2685         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
2686                 zone_page_state(zone, NR_ACTIVE_FILE);
2687
2688         /*
2689          * It's possible for there to be more file mapped pages than
2690          * accounted for by the pages on the file LRU lists because
2691          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2692          */
2693         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
2694 }
2695
2696 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
2697 static long zone_pagecache_reclaimable(struct zone *zone)
2698 {
2699         long nr_pagecache_reclaimable;
2700         long delta = 0;
2701
2702         /*
2703          * If RECLAIM_SWAP is set, then all file pages are considered
2704          * potentially reclaimable. Otherwise, we have to worry about
2705          * pages like swapcache and zone_unmapped_file_pages() provides
2706          * a better estimate
2707          */
2708         if (zone_reclaim_mode & RECLAIM_SWAP)
2709                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
2710         else
2711                 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
2712
2713         /* If we can't clean pages, remove dirty pages from consideration */
2714         if (!(zone_reclaim_mode & RECLAIM_WRITE))
2715                 delta += zone_page_state(zone, NR_FILE_DIRTY);
2716
2717         /* Watch for any possible underflows due to delta */
2718         if (unlikely(delta > nr_pagecache_reclaimable))
2719                 delta = nr_pagecache_reclaimable;
2720
2721         return nr_pagecache_reclaimable - delta;
2722 }
2723
2724 /*
2725  * Try to free up some pages from this zone through reclaim.
2726  */
2727 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2728 {
2729         /* Minimum pages needed in order to stay on node */
2730         const unsigned long nr_pages = 1 << order;
2731         struct task_struct *p = current;
2732         struct reclaim_state reclaim_state;
2733         int priority;
2734         struct scan_control sc = {
2735                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2736                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2737                 .may_swap = 1,
2738                 .nr_to_reclaim = max_t(unsigned long, nr_pages,
2739                                        SWAP_CLUSTER_MAX),
2740                 .gfp_mask = gfp_mask,
2741                 .swappiness = vm_swappiness,
2742                 .order = order,
2743         };
2744         unsigned long nr_slab_pages0, nr_slab_pages1;
2745
2746         cond_resched();
2747         /*
2748          * We need to be able to allocate from the reserves for RECLAIM_SWAP
2749          * and we also need to be able to write out pages for RECLAIM_WRITE
2750          * and RECLAIM_SWAP.
2751          */
2752         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2753         lockdep_set_current_reclaim_state(gfp_mask);
2754         reclaim_state.reclaimed_slab = 0;
2755         p->reclaim_state = &reclaim_state;
2756
2757         if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
2758                 /*
2759                  * Free memory by calling shrink zone with increasing
2760                  * priorities until we have enough memory freed.
2761                  */
2762                 priority = ZONE_RECLAIM_PRIORITY;
2763                 do {
2764                         shrink_zone(priority, zone, &sc);
2765                         priority--;
2766                 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
2767         }
2768
2769         nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2770         if (nr_slab_pages0 > zone->min_slab_pages) {
2771                 /*
2772                  * shrink_slab() does not currently allow us to determine how
2773                  * many pages were freed in this zone. So we take the current
2774                  * number of slab pages and shake the slab until it is reduced
2775                  * by the same nr_pages that we used for reclaiming unmapped
2776                  * pages.
2777                  *
2778                  * Note that shrink_slab will free memory on all zones and may
2779                  * take a long time.
2780                  */
2781                 for (;;) {
2782                         unsigned long lru_pages = zone_reclaimable_pages(zone);
2783
2784                         /* No reclaimable slab or very low memory pressure */
2785                         if (!shrink_slab(sc.nr_scanned, gfp_mask, lru_pages))
2786                                 break;
2787
2788                         /* Freed enough memory */
2789                         nr_slab_pages1 = zone_page_state(zone,
2790                                                         NR_SLAB_RECLAIMABLE);
2791                         if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
2792                                 break;
2793                 }
2794
2795                 /*
2796                  * Update nr_reclaimed by the number of slab pages we
2797                  * reclaimed from this zone.
2798                  */
2799                 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2800                 if (nr_slab_pages1 < nr_slab_pages0)
2801                         sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
2802         }
2803
2804         p->reclaim_state = NULL;
2805         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
2806         lockdep_clear_current_reclaim_state();
2807         return sc.nr_reclaimed >= nr_pages;
2808 }
2809
2810 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2811 {
2812         int node_id;
2813         int ret;
2814
2815         /*
2816          * Zone reclaim reclaims unmapped file backed pages and
2817          * slab pages if we are over the defined limits.
2818          *
2819          * A small portion of unmapped file backed pages is needed for
2820          * file I/O otherwise pages read by file I/O will be immediately
2821          * thrown out if the zone is overallocated. So we do not reclaim
2822          * if less than a specified percentage of the zone is used by
2823          * unmapped file backed pages.
2824          */
2825         if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
2826             zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
2827                 return ZONE_RECLAIM_FULL;
2828
2829         if (zone->all_unreclaimable)
2830                 return ZONE_RECLAIM_FULL;
2831
2832         /*
2833          * Do not scan if the allocation should not be delayed.
2834          */
2835         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
2836                 return ZONE_RECLAIM_NOSCAN;
2837
2838         /*
2839          * Only run zone reclaim on the local zone or on zones that do not
2840          * have associated processors. This will favor the local processor
2841          * over remote processors and spread off node memory allocations
2842          * as wide as possible.
2843          */
2844         node_id = zone_to_nid(zone);
2845         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
2846                 return ZONE_RECLAIM_NOSCAN;
2847
2848         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2849                 return ZONE_RECLAIM_NOSCAN;
2850
2851         ret = __zone_reclaim(zone, gfp_mask, order);
2852         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2853
2854         if (!ret)
2855                 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
2856
2857         return ret;
2858 }
2859 #endif
2860
2861 /*
2862  * page_evictable - test whether a page is evictable
2863  * @page: the page to test
2864  * @vma: the VMA in which the page is or will be mapped, may be NULL
2865  *
2866  * Test whether page is evictable--i.e., should be placed on active/inactive
2867  * lists vs unevictable list.  The vma argument is !NULL when called from the
2868  * fault path to determine how to instantate a new page.
2869  *
2870  * Reasons page might not be evictable:
2871  * (1) page's mapping marked unevictable
2872  * (2) page is part of an mlocked VMA
2873  *
2874  */
2875 int page_evictable(struct page *page, struct vm_area_struct *vma)
2876 {
2877
2878         if (mapping_unevictable(page_mapping(page)))
2879                 return 0;
2880
2881         if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2882                 return 0;
2883
2884         return 1;
2885 }
2886
2887 /**
2888  * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2889  * @page: page to check evictability and move to appropriate lru list
2890  * @zone: zone page is in
2891  *
2892  * Checks a page for evictability and moves the page to the appropriate
2893  * zone lru list.
2894  *
2895  * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2896  * have PageUnevictable set.
2897  */
2898 static void check_move_unevictable_page(struct page *page, struct zone *zone)
2899 {
2900         VM_BUG_ON(PageActive(page));
2901
2902 retry:
2903         ClearPageUnevictable(page);
2904         if (page_evictable(page, NULL)) {
2905                 enum lru_list l = page_lru_base_type(page);
2906
2907                 __dec_zone_state(zone, NR_UNEVICTABLE);
2908                 list_move(&page->lru, &zone->lru[l].list);
2909                 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
2910                 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
2911                 __count_vm_event(UNEVICTABLE_PGRESCUED);
2912         } else {
2913                 /*
2914                  * rotate unevictable list
2915                  */
2916                 SetPageUnevictable(page);
2917                 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
2918                 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
2919                 if (page_evictable(page, NULL))
2920                         goto retry;
2921         }
2922 }
2923
2924 /**
2925  * scan_mapping_unevictable_pages - scan an address space for evictable pages
2926  * @mapping: struct address_space to scan for evictable pages
2927  *
2928  * Scan all pages in mapping.  Check unevictable pages for
2929  * evictability and move them to the appropriate zone lru list.
2930  */
2931 void scan_mapping_unevictable_pages(struct address_space *mapping)
2932 {
2933         pgoff_t next = 0;
2934         pgoff_t end   = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
2935                          PAGE_CACHE_SHIFT;
2936         struct zone *zone;
2937         struct pagevec pvec;
2938
2939         if (mapping->nrpages == 0)
2940                 return;
2941
2942         pagevec_init(&pvec, 0);
2943         while (next < end &&
2944                 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
2945                 int i;
2946                 int pg_scanned = 0;
2947
2948                 zone = NULL;
2949
2950                 for (i = 0; i < pagevec_count(&pvec); i++) {
2951                         struct page *page = pvec.pages[i];
2952                         pgoff_t page_index = page->index;
2953                         struct zone *pagezone = page_zone(page);
2954
2955                         pg_scanned++;
2956                         if (page_index > next)
2957                                 next = page_index;
2958                         next++;
2959
2960                         if (pagezone != zone) {
2961                                 if (zone)
2962                                         spin_unlock_irq(&zone->lru_lock);
2963                                 zone = pagezone;
2964                                 spin_lock_irq(&zone->lru_lock);
2965                         }
2966
2967                         if (PageLRU(page) && PageUnevictable(page))
2968                                 check_move_unevictable_page(page, zone);
2969                 }
2970                 if (zone)
2971                         spin_unlock_irq(&zone->lru_lock);
2972                 pagevec_release(&pvec);
2973
2974                 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
2975         }
2976
2977 }
2978
2979 /**
2980  * scan_zone_unevictable_pages - check unevictable list for evictable pages
2981  * @zone - zone of which to scan the unevictable list
2982  *
2983  * Scan @zone's unevictable LRU lists to check for pages that have become
2984  * evictable.  Move those that have to @zone's inactive list where they
2985  * become candidates for reclaim, unless shrink_inactive_zone() decides
2986  * to reactivate them.  Pages that are still unevictable are rotated
2987  * back onto @zone's unevictable list.
2988  */
2989 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
2990 static void scan_zone_unevictable_pages(struct zone *zone)
2991 {
2992         struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
2993         unsigned long scan;
2994         unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
2995
2996         while (nr_to_scan > 0) {
2997                 unsigned long batch_size = min(nr_to_scan,
2998                                                 SCAN_UNEVICTABLE_BATCH_SIZE);
2999
3000                 spin_lock_irq(&zone->lru_lock);
3001                 for (scan = 0;  scan < batch_size; scan++) {
3002                         struct page *page = lru_to_page(l_unevictable);
3003
3004                         if (!trylock_page(page))
3005                                 continue;
3006
3007                         prefetchw_prev_lru_page(page, l_unevictable, flags);
3008
3009                         if (likely(PageLRU(page) && PageUnevictable(page)))
3010                                 check_move_unevictable_page(page, zone);
3011
3012                         unlock_page(page);
3013                 }
3014                 spin_unlock_irq(&zone->lru_lock);
3015
3016                 nr_to_scan -= batch_size;
3017         }
3018 }
3019
3020
3021 /**
3022  * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
3023  *
3024  * A really big hammer:  scan all zones' unevictable LRU lists to check for
3025  * pages that have become evictable.  Move those back to the zones'
3026  * inactive list where they become candidates for reclaim.
3027  * This occurs when, e.g., we have unswappable pages on the unevictable lists,
3028  * and we add swap to the system.  As such, it runs in the context of a task
3029  * that has possibly/probably made some previously unevictable pages
3030  * evictable.
3031  */
3032 static void scan_all_zones_unevictable_pages(void)
3033 {
3034         struct zone *zone;
3035
3036         for_each_zone(zone) {
3037                 scan_zone_unevictable_pages(zone);
3038         }
3039 }
3040
3041 /*
3042  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
3043  * all nodes' unevictable lists for evictable pages
3044  */
3045 unsigned long scan_unevictable_pages;
3046
3047 int scan_unevictable_handler(struct ctl_table *table, int write,
3048                            void __user *buffer,
3049                            size_t *length, loff_t *ppos)
3050 {
3051         proc_doulongvec_minmax(table, write, buffer, length, ppos);
3052
3053         if (write && *(unsigned long *)table->data)
3054                 scan_all_zones_unevictable_pages();
3055
3056         scan_unevictable_pages = 0;
3057         return 0;
3058 }
3059
3060 #ifdef CONFIG_NUMA
3061 /*
3062  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
3063  * a specified node's per zone unevictable lists for evictable pages.
3064  */
3065
3066 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
3067                                           struct sysdev_attribute *attr,
3068                                           char *buf)
3069 {
3070         return sprintf(buf, "0\n");     /* always zero; should fit... */
3071 }
3072
3073 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
3074                                            struct sysdev_attribute *attr,
3075                                         const char *buf, size_t count)
3076 {
3077         struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
3078         struct zone *zone;
3079         unsigned long res;
3080         unsigned long req = strict_strtoul(buf, 10, &res);
3081
3082         if (!req)
3083                 return 1;       /* zero is no-op */
3084
3085         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
3086                 if (!populated_zone(zone))
3087                         continue;
3088                 scan_zone_unevictable_pages(zone);
3089         }
3090         return 1;
3091 }
3092
3093
3094 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3095                         read_scan_unevictable_node,
3096                         write_scan_unevictable_node);
3097
3098 int scan_unevictable_register_node(struct node *node)
3099 {
3100         return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
3101 }
3102
3103 void scan_unevictable_unregister_node(struct node *node)
3104 {
3105         sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
3106 }
3107 #endif